10 Ways to Hide Data for Microsoft Copilot and Copilot for M365

10 Ways to Hide Data for Microsoft Copilot and Copilot for M365

Microsoft Copilot and Copilot for M365 are powerful AI tools designed to enhance productivity by generating context-aware suggestions. However, ensuring the privacy and security of sensitive data while using these tools is crucial. Here are 10 strategies to hide and protect data effectively when leveraging Microsoft Copilot and Copilot for M365.


1. Leverage Sensitivity Labels in Microsoft 365

Sensitivity labels allow you to classify and protect content based on its level of confidentiality:

  • Mark documents as Confidential or Highly Confidential to restrict access.
  • Prevent Copilot from analyzing documents with specific sensitivity labels.
  • Set rules for email handling, such as encryption or blocking unauthorized forwarding.

This ensures sensitive data is flagged and controlled automatically.


2. Disable Copilot for Specific Content

For particularly sensitive documents or emails, disable Copilot suggestions entirely:

  • In Microsoft 365, turn off Copilot features for selected files or emails.
  • Limit Copilot's access in administrative settings to ensure specific categories of content are excluded.

By doing so, you control where Copilot operates.


3. Use Redacted Copies of Documents

Before sharing or processing sensitive documents, create redacted versions:

  • Replace sensitive sections with placeholders like [REDACTED].
  • Use tools such as Microsoft Word's Redact feature to black out confidential information.

This ensures that Copilot interacts with sanitized content.


4. Control Access with Permissions

Utilize Microsoft 365’s role-based access controls (RBAC) to limit who can view or edit sensitive files:

  • Share documents with specific individuals or groups.
  • Use permissions to restrict editing, copying, or printing of sensitive documents.

This limits the exposure of sensitive data in shared environments.


5. Mask Data in Documents

Replace sensitive information with dummy data or anonymized placeholders:

  • Replace real names, addresses, or account numbers with generic terms like NAME_PLACEHOLDER or 123-456-7890.
  • Use Microsoft Excel's built-in anonymization tools for spreadsheets.

Masked data allows Copilot to work effectively without risking privacy.


6. Use Information Protection Policies

Microsoft Purview Information Protection (MIP) allows you to apply protection policies across your organization:

  • Automatically encrypt emails and files containing sensitive data.
  • Set policies to block Copilot from processing data marked as confidential.
  • Monitor and audit how sensitive content is accessed and shared.

This provides enterprise-level data security.


7. Implement Copilot Governance Settings

Configure Microsoft Copilot's governance features to control its behavior:

  • Limit Copilot's contextual analysis scope.
  • Disable Copilot in areas where sensitive data is handled frequently, such as HR or legal departments.

Governance settings let you tailor Copilot usage to your organization’s security requirements.


8. Encrypt Emails and Files

Encryption ensures sensitive data remains secure even if accessed by unauthorized parties:

  • Use Microsoft 365’s built-in encryption for documents and emails.
  • Apply end-to-end encryption for particularly sensitive communication.

Encryption ensures Copilot cannot expose encrypted content inadvertently.


9. Avoid Including Sensitive Data in Prompts

When using Copilot, avoid inputting sensitive data directly in prompts or context:

  • Instead of typing an actual account number, use XXXX-XXXX-XXXX-XXXX.
  • Refrain from including real client names, proprietary formulas, or confidential terms in instructions to Copilot.

This reduces the risk of sensitive data being exposed in generated outputs.


10. Educate Teams on Safe Practices

Training your team is vital to ensure secure Copilot usage:

  • Teach employees to identify and label sensitive content correctly.
  • Share best practices for using placeholders or redacted versions when engaging with Copilot.
  • Conduct regular reviews to ensure adherence to privacy and security policies.

A knowledgeable team reduces the risk of accidental data exposure.


Final Thoughts

Microsoft Copilot and Copilot for M365 are valuable tools for enhancing productivity, but they must be used carefully to protect sensitive data. By implementing these 10 strategies, ranging from encryption and redaction to access controls and education, you can safeguard your organization's information while making the most of AI-driven tools.

Manjish Deb (CISM, CIPM, CEH, MCIPA, MCASE, SCM-DLP)

"Cybersecurity Maverick | Guardian of Data Realms | Privacy Crusader | Pioneering AI Wizardry for Tomorrow’s Security"

1 周

Hi Marcel Broschk, Thanks for sharing this article. However, I haven't had a positive experience with MS Purview's controls and monitoring mechanism on Co-Pilot's activities: 1) DLP Policies available for Co-Pilot monitoring is limited and doesn’t offer us features related to content based inspection of the uploads and queries raised by users, only SL based controls which stops processing the content on a Webchat. This feature too isn’t working, and we have an open support case with MS Support. 2)No definite way to ring fence or an ability to create an information barrier between Internal (Work) and External (Web) Co-Pilot. apart from option 1. The other 2 options are: ?Encrypting the document (Collaboration Nightmare) ?Disable Web-Search for Co-pilot (counter-productive) 3) MS Purview controls for MS Co-Pilot is only available in Preview mode and as of now available for Co-Pilot Web Chat mode. For all non-Web Chat mode Co-Pilot Interactions, MS Purview has no Monitoring or Controls. Moreover, as this feature is available in Preview mode, we can’t initiate any DLP Policies or Rules which could work in Simulation Mode.? These are my 2 cents based on our testing and observations.

回复
Gareth Naylor

Groupwide IT Strategist/Architect at Uniper

2 个月

Thanks! Can you point me in the direction of more information on the Word Redact tool mentioned in #3 please?

Sebastian Holst

Post Microsoft | Product Leadership, Strategy, and GTM

2 个月

All of these are excellent tactics to ensure data is excluded either from consumption or use by Copilot. It is worth noting (IMO) that some of these are specific to Copilot and will not impede other tools like chatgpt or potentially specialized LLMs. This is not a criticism, just pointing out that if these controls are seen as essential, they must be paired with prohibitions of other llms. Lastly, I would point out that this is a special case of Large Language Model Optimization (LLMO). Restricting access goes along with ensuring clarity. In the anonymization example above, u want to be sure that anonymized data is not seen as real in far removed associations inside the model. Or, for example, Contoso is not a real company and should not be considered in prompts that ask about common client usage patterns. Generating (or marking up) content specifically for LLMO is, IMO, something to consider.

really good points, but requires massive manual work, before it works and no guarantee that people will put in all the hours to do it. In addition in a normal IT setup, only roughly 10% of the data support sensitivity labels. e.g. All existing e-mails in your exchange online or on prem can get a sensitivity label.

Andreas Sieber

Make it! Right, but always with fun!

2 个月

Impressive, Marcel ??

要查看或添加评论,请登录

Marcel Broschk的更多文章

社区洞察

其他会员也浏览了